Datalab’s Marker 2 Outperforms Rivals in Speed and Accuracy on AI2 OCR Benchmark

Datalab has released Marker 2, a complete rewrite of its open-source document conversion pipeline. The tool, which converts PDFs, images, and office documents into markdown, JSON, or HTML, has been rebuilt around new core components including the Surya OCR 2 engine and a faster pdftext extractor. According to benchmark results shared by the company, Marker 2 achieves a leading balance of accuracy and speed against several competing systems.

Key takeaways

• Marker 2’s ‘balanced’ mode scores 76.0% overall on the third-party olmOCR-bench from Allen AI, outperforming MinerU (72.7%), Docling (50.3%), and LiteParse (22.4%).
• Throughput is a major differentiator: Marker 2 processes 2.9 pages per second, over 5 times faster than MinerU’s pipeline backend at 0.54 pages per second.
• The system introduces three conversion paths: a high-quality ‘balanced’ mode for GPU, a ‘fast’ CPU/GPU hybrid mode scoring 66.6%, and a CPU-only ‘–disable_ocr’ mode achieving 43.6% accuracy at 23.7 pages per second.
• Architectural changes enable high concurrency, with many CPU workers sharing a single Surya inference server, allowing throughput to scale with server capacity.
• While MinerU’s VLM backend was not included in the comparison, Datalab notes it scores higher than MinerU’s pipeline backend, representing a different, full-page-VLM approach.

Performance and competitive analysis

The primary benchmark cited is olmOCR-bench from AI2 (Allen Institute for AI), which uses 1,403 PDFs and approximately 8,400 tests covering math, tables, reading order, and more. Datalab’s reported figures are from their own reproducible runs using an open harness in the Marker repository that includes competitor runners. Against MinerU’s pipeline backend—the closest architectural match—Marker 2’s balanced mode leads in overall score (76.0% vs. 72.7%) and is effectively tied on born-digital PDFs (83.5% vs. 83.3%). The throughput gap is significant, with Marker achieving 2.9 pages/second versus MinerU’s 0.54 pages/second. Marker’s ‘fast’ mode pushes speed further to 7.4 pages/second, though with a lower accuracy score of 66.6%.

Compared to Docling, Marker 2 shows a wider accuracy margin, leading 76.0% to 50.3% overall and 83.5% to 64.0% on born-digital documents, while also being faster (2.9 vs. 2.1 pages/second). Datalab notes Docling’s strengths lie in its governance—being MIT-licensed and part of the LF AI & Data Foundation—and its support for a broader range of input formats including audio and email. LiteParse, a Rust-based parser, operates on a different axis, trading all layout analysis for extreme speed; with OCR disabled, it processes 1,721 pages/second but scores only 20.4% on the benchmark. Marker’s CPU-only mode more than doubles that score by including a lightweight 20M-parameter layout model to recover document structure.

Why it matters

The release of Marker 2 highlights the ongoing evolution of document intelligence pipelines, where the trade-off between accuracy, speed, and cost is paramount for practical deployment. By offering multiple conversion paths, the tool provides flexibility for different use cases, from high-quality GPU processing to efficient CPU-only extraction. The significant throughput advantage over competitors like MinerU, while maintaining higher accuracy, could lower processing costs and latency for organizations handling large document volumes. Furthermore, Datalab’s emphasis on Marker as a pipeline distinct from a full-page Vision Language Model (VLM) clarifies the architectural choices available to AI teams, who must decide between integrated systems and specialized, modular tools for their specific document processing needs.